Real estate is one of the most data-rich industries on paper and one of the most data-messy industries in practice. Every transaction involves people, places, funding, paperwork, and time-sensitive decisions. The stakes are high because a “minor” percentage error translates into a massive amount of money. The regulations are tight because housing, lending, and advertising are all pretty tightly regulated everywhere. And workflows in real estate are riddled with edge cases because no two properties are the same, and markets move fast.
That combination is precisely what enables real estate to benefit from AI, but also what causes many “AI projects” to flounder. You don’t get to add a chatbot to a broken process and win. You win when AI is connected to a measurable business outcome, fueled by reliable data, wrapped in governance, and embedded in the systems people already use.
This is a comprehensive founder’s guide to the best, most impactful AI today in real estate. It is written for Western product leaders who seek practical implementation details instead of hype. everything, including “what,” “how,” “what can go wrong,” and regulations and compliance, and financial and operational realities that a serious team must plan for.
To keep it grounded, you’ll find practical real-world applications from popular platforms such as Zillow’s Neural Zestimate, which employs deep learning for massive-scale home valuation, to mortgage-side collateral risk assessment tooling like Fannie Mae’s Collateral Underwriter, touted as an automated appraisal risk assessment tool that can assess risk on most appraisals it receives. Learn too how regulators are increasingly actively defining what “credible and safe” AI means in real estate valuation through the U.S. interagency Automated Valuation Models (AVM) final rule and how housing regulators are grappling with algorithmic discrimination risks in tenant screening and housing advertising.
The bulk of AI-powered real estate value falls into three buckets.
The first being revenue and conversion. This can be lead scoring, speed-to-lead, appointment booking, and enhancing agent or leasing workflows so that fewer leads die in silence.
The second is an issue of valuation and risk. These include automated valuation models (AVMs), condition assessment, appraisal quality control, underwriting signals, and portfolio risk monitoring.
Third is experience and operations. These are virtual tours, digital twins, support automation, maintenance, energy optimization, and operational intelligence that trim cost-to-serve.
The ten solutions below are the “most common” themes you see across brokerages, listing portals, lenders, property managers, REITs, developers, and proptech platforms. In reality, a real estate firm typically requires three or four of these partnering with each other, since each enhances the data and results of the others.
Lead scoring for real estate seems obvious, but the details are more important than in most industries. Real estate leads are high intent, but they are also highly perishable. They’re also noisy. Many are “curious clicks.” Many are duplicate household inquiries over multiple devices. A lot are mismatched to the local agent’s inventory or area of service. And plenty go cold just because nobody got back to them quickly enough.
One classic bit of speed-to-lead evidence comes from the Harvard Business Review study on online lead response behavior, which found that many companies are responding far too late and that leads go cold fast. Real estate compounds that, as homebuyers and renters, can be sniffing around half a dozen agents or listings in a matter of minutes.
A genuine lead scoring system isn’t “a model that predicts conversion.” That is just the core. The true engine also determines the next step: which agent should be assigned the lead, what message should be sent out, when a human should intervene, and what follow-up sequence makes sense.
Top-performing implementations treat scoring as a routing/prioritization engine and not merely a dashboard metric. You typically start with a baseline model that estimates the likelihood of meaningful engagement, such as a completed call, scheduled viewing, submitted application, or mortgage pre-qual. Now add some constraints: service area match, language match, availability match, deal size, specialty (luxury, first-time buyer, or commercial), and fairness guardrails (e.g., not using protected traits, not using proxy features that recreate protected traits).
The data labor is usually the hardest. A serious lead model incorporates a single view of inquiry behavior, listing engagement, channel source, recency, and frequency of patterns, and operational outcomes. A lot of teams fall into the trap of only training on “closed deals.” That’s too slow and too biased. In real estate, the closing of the feedback loop from inquiry to close can be months. You want intermediate labels that happen faster, like “connected call,” “attended showing,” “submitted docs,” “application completed,” or “pre-approval started.” That’s how you get a model that learns in weeks, not quarters.
Detail of implementation that matters: deduplication and resolution of identity. A “lead” is frequently the same household on multiple devices and emails. If your data treats them as two different people, you’ll have inconsistent scoring and bad conversion attribution, and your agents will lose trust. Most realtor orgs will create a lightweight identity graph, which connects email, phone, cookie/device, and CRM contact IDs into a “household entity,” with well-defined rules for merging and for audit.
The choice of model is usually nothing exotic. Gradient boosted trees are prevalent due to their robust performance on mixed structured features and fast training. Calibration is the more interesting part, since we want good decisions, which depend on good probability. Calibration is important when you apply threshold rules such as “call within 2 minutes if score > X.”
The “why this fails” model is also applicable. Teams develop a score that performs well in offline AUC but never iterate on the workflow. Agents still choose leads based on gut feel. Or the CRM is still a mess. Or the “routing rules” battle local team politics. Or no one defines a quantifiable win beyond “better ML.”
An ideal mature result would be: the system facilitates higher contact and appointment rates, shorter lead response times, greater conversion per agent hour, and it does not drive complaints or propagate unfair filtering. You measure success in operational metrics, not model metrics.
It is based on priority for lead scoring. Conversational AI secures the lead before it goes cold.
In real estate, a lot of queries come after hours. A lot of renters message late. Submit forms on weekends. Responding to buyers outside of business hours is an obvious edge you’re giving away to your competitors. This is where conversational AI makes sense, but only if it’s done as a transaction workflow and not as “chat.”
The easiest variant is an AI assistant that instantly answers queries, asks a brief series of qualifying questions, and schedules the next step. The “next step” varies by business: a call with an agent, an appointment for a showing, a self-tour link, a rental application, or a mortgage pre-qualification flow.
The most important product decision is the behavior of the agent when it doesn’t know. A real estate assistant can not make things up about availability, pricing, fees, school zones, HOA rules, or legal terms. The secure pattern is to have an assistant grounded on trusted data sources, and when the data is not known, ask the user to confirm or escalate to a human.
In reality, this is a question of orchestration. The assistant must have access to listings and availability calendars, CRM records, call center schedules, showing management tools, and, in some cases, payment and identity verifications. It should also log every interaction back into the CRM, because otherwise your future lead scoring and follow-up logic goes blind.
The “serious” edition has a voice. Many leads call, not chat. Use speech-to-text and intent recognition to collect structured data during calls, to summarize conversations, and extract action items. This is not just about productivity. It is training data. When you determine why leads drop off, you understand what inventory, pricing, or messaging problems are killing deals.
Risk and compliance: If you do business in a regulated market, your conversational AI must be compliant with fair housing rules and advertising rules. HUD has flagged fair housing concerns in AI-based tenant screening and advertising in the U.S. Even if your chatbot is “just answering questions,” it can still be part of a discriminatory game if it nudges different users slightly differently based on signals that correlate with protected traits.
A common design pattern is to treat the assistant as a “process guide” with a known policy. The model can paraphrase and chat, but the real decisions (e.g., who is eligible, what the screening thresholds are, what rules apply to ads) must be transparent, logged, and auditable.
Success metrics here are very much operational: response time, contact rate, booked appointment rate, after-hours captured demand, no-show reduction, and agent time saved. You may also want to measure the rate of escalation to a human and user satisfaction, because if your assistant irritates users, conversion will drop even if your response time improves.
Real estate valuation is the eye of the hurricane in real estate AI. It’s also a source of a great deal of confusion when it comes to AI.
There is a difference between an AVM for marketing (“How much is this home worth?”) and an AVM for lending (“What is credible collateral value under policy and regulation?”). There is also a difference between a point estimate and an estimate with uncertainty bounds and explainable drivers.
Among the most well-known consumer valuation systems is Zillow’s Zestimate. Zillow has explained how it enhanced its methodology using deep learning in the “Neural Zestimate,” which applies to over 100 million homes and updates regularly with market signals. Zillow also explains how past models consisted of many regional models and multi-step pipelines, whereas the neural model consolidates maintainability and accuracy. This is useful information, not because you want to replicate Zillow, but because it reveals the real limitations: national scale, diverse property types, volatile markets, incomplete data, and the requirement for continual iteration.
On the loan side, U.S. regulators have specifically targeted the reliability and integrity of AVMs. In 2024, several agencies issued a final rule to establish quality control requirements for AVMs utilized by mortgage originators and offers in the secondary market. The existence of this rule should serve as a very strong signal: AVMs are not “just analytics.” They are considered decision-critical models and thus require governance, controls, and monitoring.
If you’re building AVMs for a proptech product, the most important technical point is not the model type. It is the data definition “truth,” and how you manage drift.
The “true value” in an appraisal is not the transaction price. Deals could involve concessions, distressed sales, transfers off the market, or stale information. Listings may be strategically priced. Appraisals differ. If you train without controlling for these, your model learns noise.
A sophisticated valuation model has several labels depending on the use case: sales price for sold properties, appraisal values where applicable, rent contract values for rent estimates, and repeat-sale indices for market movement. It also performs time-aware validation to prevent future data from leaking into training.
Modeling techniques vary with maturity. Many systems begin with hedonic regression plus neighborhood components. Then they move to gradient boosting or neural networks. Then they add text and image signals. Images are important because condition is important, and condition is often not captured in structured data. Recent valuation literature finds that traditional housing characteristics can be augmented with visual features extracted from Street View and aerial imagery to enhance price prediction accuracy, which is one of the reasons why computer vision is finding more and more applications in valuation processes.
Operationally, the best AVMs do not provide just a number. They provide a number, a confidence interval, explanations of drivers, and a selection of comparable properties, with rules on how comps are selected. Even if you use deep learning, you still need “comps logic,” because valuation is audited by humans.
Where this goes wrong: teams optimize for low average error, ignoring tail risk. In real estate, you want to know where the model is wrong, not just how often it is wrong. If the model tends to be wrong for certain neighborhoods, certain property types, or certain price tiers, that becomes a business risk and reputational risk, and potentially a compliance risk, in some contexts.
Good deployment features include evaluation on the segment level, drift detection, and a defined “human override” process. You also need a policy for when not to use the model — in areas that are very unique, in thinly traded neighborhoods, or in volatile markets.
Whilst AVMs estimate value, appraisal QC processes estimate risk within the process of valuation itself.
That’s important because many real estate companies require more than just an estimated value. They want to know if an appraisal is credible, if the comps are reasonable, and if the valuation is an outlier when compared to data.
Fannie Mae’s Collateral Underwriter (CU) is perhaps the best-known example. Fannie Mae characterizes it as a proprietary automated appraisal risk assessment model that measures risk for most appraisals received for single-family loans it acquires. The key concept here is that the system is taking appraisal data and comparing it to massive comps and property data sets to identify valuation risk, not replacing appraisers.
To a product leader building in this space, the insight is that you can generate “AI value” without making the final decision. You generate value through better quality control, shorter cycle time, and by focusing human attention on the highest risk cases.
The implementation pattern typically goes like this: You consume appraisal reports, you normalize the structured fields, you extract additional signals from text narratives, you compute comp similarity metrics, and you compare the appraisal value to independent benchmarks and to local market distributions. Then you get a risk score and a “reason report.”
This is also a place where explainability isn’t optional. Underwriters and reviewers have to be able to see why it’s flagged. Was the comp selection unusual? Was the adjustment size extreme? Was the subject property lacking in the report of key features? Was the market thin?
This category lends itself well to AI as it is a combination of rules and models. Rules represent policy. Models identify anomalies. Text extraction extracts structured data from unstructured fields. A “pure deep learning” solution is rarely allowed because the business needs to be traceable.
If you’re outsourcing this build, you want a team that knows both ML and mortgage/appraisal workflows. A lot of technically capable teams fail at this because they treat appraisal reports as just any other document, but they actually have domain structure, terminology, and policy constraints.
Condition is among the largest unseen variables in real estate. Two houses that are the same size, in the same neighborhood, and on the same street can be worth considerably different amounts due to the interior condition, quality of layout, amount of natural light, maintenance on the home, and renovations.
Conventional structured datasets capture some, but not all, of this. That is the reason why computer vision has become the core of modern real estate AI.
There are two main modes of application.
The first is appraisal enrichment. You apply real property photos, street view, and aerial view images to extract features that correlate with value. Street view and satellite images have been shown in academic work to capture “urban qualities” and to add predictive power to house price estimation when used in conjunction with traditional attributes. Other research further builds this into interior and exterior visual features to predict house prices, also highlighting the importance of vision signals in valuation tasks
The second is operations automation. You apply your vision to identify problems in inspection images, develop standardized condition scores, and reduce the amount of manual review. For property managers, that could be a matter of identifying water damage, mold risk factors, appliance condition, or safety concerns. For insurers, it might be roof condition or wildfire defensible space indicators, depending on the region.
A closely related, but distinct problem is listing quality. Real estate marketplaces are dependent on listing conversion. AI can evaluate photo quality, identify missing photo categories (kitchen, bathroom, exterior), verify images correspond to property type, and detect misleading images. This is not glamorous, but it boosts trust and conversion.
Important details of the implementation are the labeling policy. If you attempt to construct a “condition score” without a well-defined label, it becomes subjective noise. Top teams tie their condition labels to outcomes such as estimates of renovation costs, days on market, price reductions, inspection findings, or (when available) appraisal condition ratings. They also use “style” and “condition,” since an outdated “style” is not necessarily a bad “condition.”
One big risk is fairness and proxies. Neighborhood images are correlated with income levels. If you apply neighborhood imagery features in models that determine who gets to see what, or who gets prioritized, you may inadvertently introduce discriminatory outcomes into your system. Which is why AI governance frameworks such as NIST’s AI Risk Management Framework encourage consideration of context, impacts, and monitoring throughout the system lifecycle.
Real estate requires a lot of paperwork. Transactions flow through leases, addenda, title docs, HOA rules, inspection reports, appraisal reports, mortgage disclosures, insurance certificates, and compliance documents.
AI document intelligence is one of the quickest ways to reduce cycle time and errors, as it addresses a glaring bottleneck: people copying fields from PDFs into systems and back.
The real-life equivalent of this solution is not “OCR.” It is an extraction and validation pipeline. First, you classify the document type. Based on that, you extract key fields using a mixture of layout-aware models, rules, and, in some cases, human-in-the-loop review for low-confidence decisions. And then you validate those fields with other sources. Then you write the structured output back into the system of record, along with an audit log of what was extracted and why.
For instance, with leasing, you can pull the tenant name, rent amount, lease term, renewal options, deposit, fee schedule, and any special terms. In mortgage processing, you can capture income, liabilities, employment, and appraisal fields. In commercial real estate, you can extract rent rolls and CAM clauses and then normalize them.
The “implementation trap” is thinking of this as just a generic LLM task. It’s not. Transaction documents need to be deterministic and traceable. A robust design employs a model as an assistant, but the final extracted outputs are subjected to a set of deterministic validation rules. Should a rent be extracted, you check the currency, the format, and that it is the same value as in the schedule in the document. When an address is extracted, you normalize and geocode it. When a clause is extracted, you keep a reference to the exact paragraph and page.
This is also true for privacy and security by design that matters. The documents include sensitive personal information. A lot of companies require strong access controls, encryption at rest, and retention rules. If you do business in multiple geographic regions, you need to build for the GDPR and other privacy frameworks.
A good product call, should we build this in-house or buy specialized tooling? If you can standardize on documents and have a high volume, buying can be faster. If documents are messy, multi-language, and tightly integrated with your own proprietary workflow, then building can be worth it. The hybrid scenario is very common: get a vendor for your OCR/layout extraction part, but develop your domain validation/workflow integration yourself.
Tenant screening is a high-stakes application of AI, since it involves access to housing. It is also an area where regulators are looking closely.
HUD has published guidance on the application of the Fair Housing Act to tenant screening and housing advertising in the context of AI and similar algorithm-based tools, and underscores that such tools could inadvertently cause discriminatory effects if they are poorly designed or used without appropriate safeguards. Even beyond the United States, analogous principles exist under anti-discrimination law and consumer protection.
In terms of product, the purpose of screening is often described as “reduce default risk.” But the end-to-end goal is “to manage risk effectively within an accurate, explainable, contestable and fair process.”
Separates the best implementations into three things.
First, you need to do the identity and fraud checks. Is the applicant genuine? Are the documents genuine? Any signs of synthetic identity activity?
Second, eligibility checks. Is basic minimum criteria satisfied according to the clear policy? It is rule-based.
Third, risk scoring. For an eligible applicant, what is the estimated likelihood of default or other poor tenancy outcomes, within the legal and ethical bounds of the market?
This distinction is important because a lot of legal and fairness questions arise when teams combine them into a single, opaque score. If the model combines signals of eligibility, risk, and fraud, it is difficult to explain and audit.
A good filtering system also needs a dispute and correction loop. A number of screening datasets are flawed. People have multiple credit files. Addresses are incorrect. If you don’t provide applicants with a means to contest decisions and correct data, you generate harm and operational load in the form of complaints and legal exposure.
This is also where “feature selection” is not a purely technical choice. A lot of features are proxies for protected characteristics. You don’t have to put in race or religion to generate something like that — you can tease out similar segments by region, levels of education, work history patterns, or whatever signals that are correlated. That’s why governance frameworks and domain legal review need to be baked in to the build, not tacked on as an afterthought.
If you’re outsourcing that, you want a team that’s going to help you not just build a classifier, but define the boundaries of the policy, build audit logging, and then design the user experience for transparency and recourse.
Real estate fraud is not just one problem. It’s a whole series of problems. There are rental scams, fake listings, identity fraud, payment fraud, forged documents, title fraud, and collusion in all of those, and AI can help – but only if you really articulate the threat model.
For marketplaces, one of the highest impact applications is fake listing detection. AI can identify duplicate images used in unrelated listings, suspicious pricing patterns, location metadata that does not match, as well as textual patterns typically associated with known scams. When it comes to rental websites, AI can flag unusual application trends and signals at the device level that suggest synthetic identities or that bots are being employed.
For transaction processing, document forgery detection counts. This may involve inspecting metadata, identifying image manipulation artifacts, verifying that numbers are consistent across pages, and comparing extracted fields against external databases.
The reason this approach tends to be effective is that fraud detection is essentially an anomaly detection task, and AI is good at triage. You don’t have to be perfectly right. You have to lessen the burden by bringing up the highest-risk items for a human review, and continually learning from the outcomes.
The greatest design risk here is over-blocking. In housing situations, false positives are expensive because they can keep bona fide people out. That’s why you need a tiered response. For low-confidence anomalies, you request additional verification. For high-confidence anomalies, you need block and log. And you want to be monitoring for bias, because your fraud models can inadvertently end up over-flagging certain groups if your training data reflects historical enforcement bias.
Rental revenue management is arguably the most contentious application of AI in real estate today, as it affects affordability, competition, and antitrust.
That the business value potential is significant is clear. Rental operators want improved occupancy, improved renewal strategy, and improved pricing levers. AI can predict demand, churn, and support concession strategy and pricing recommendations, taking into account inventory, seasonality, and local market factors.
But this is also a realm in which the distinction between “optimization” and “coordination” takes on legal significance.
The U.S. Department of Justice filed a lawsuit against RealPage in August 2024, accusing it of an algorithmic pricing scheme that harmed renters, and outlined how algorithmic tools could help enable alignment of rents utilizing shared competitively sensitive data. Late 2025 reporting discussed settlement negotiations and limitations on what data can be used and how features may need to change. Your product doesn’t even have to be that product to take the lesson: pricing AI in housing is a regulated-risk niche, and you need to design for that.
If you’re developing rental pricing AI today, you should assume that governance and auditability are part of that. Your system should be able to demonstrate what data it used, how it arrived at its recommendations, and that it is not using illegal forms of competitor data sharing. You also need internal policy controls on what “market data” is permissible, how fresh it is allowed to be, and if it is able to come from sources that cause you to have antitrust exposure.
From a product perspective, one way to play it safer is to position Pricing AI more as an internal decision support tool leveraging your own universe of inventory and performance data, along with publicly available market signals, and not any sort of sensitive competitor feed. Another safe pattern to focus on, at least depending on the legal environment and counsel advice, is demand forecasting and concession optimization rather than direct rent recommendation.
And there’s a trust aspect, even when it’s legal. Opaque pricing will be closely examined by tenants and regulators. If you can’t tell people why prices are moving, you’ll have reputational risk. And so explainability is non-negotiable. It’s part of the product.
This is the most consumer-facing “AI for real estate” bucket, and it has grown a lot beyond basic 360 tours.
The first tier is demand. Buyers and renters are being conditioned to expect rich media. The National Association of Realtors has also reported that home buyers’ agents consider photos to be very important, and that videos and virtual tours are also important to clients, suggesting that virtual tours are among what modern buyers want in listings. This is not a conversion guarantee, but it is an expectation signal.
The second layer is the production. Tools like Matterport helped bring mainstream the idea of producing immersive 3D renders of spaces — often referred to as “digital twins” — that can be leveraged for marketing and running business use cases. The key point is that when you have a 3D model, you can do more than just show it. You can measure it, create floor plans, plan renovations, run facilities, and add data overlays.
The third layer is the AI enhancement. AI can assist with tour creation from fewer captures, enhance image quality, produce object labels, and auto-generate marketing materials. AI virtual staging is capable of creating furnished versions of empty homes and renovation previews, which can help close the “imagination gap” for buyers. But you have to deal with ethics and disclosure. In many markets, you have to disclose if the images are virtually staged when you show staged images. It’s a trust and compliance thing.
From an execution point of view, a virtual tour is more than just hosting media. It is also performance engineering. Tours need to load rapidly on mobile. They must function under a variety of network conditions. They need to plug into listing pages, crm and analytics. They need to track so you can see what users looked at, and where they dropped off, because that informs marketing decisions.
Among the most overlooked opportunities is the use of tour engagement data as an intent signal. If a user examines the kitchen and bedroom, comes back to the tour several times, and shares it, these are strong signals for lead scoring. This is how the solutions integrate: media analytics augments conversion systems.
The most common failure mode is designing tours that look good but don’t move the needle on conversion. This is the result of either not measuring impact on the funnel or when the tour experience itself is so heavy that it slows down the page, hurts SEO, and causes people to leave. A proper rollout has A/B testing and performance budgets.
If you only want to take away one thing, let it be this: for real estate, the quality of the AI is mostly the quality of the data, plus the design of the workflows, plus how it’s governed. The model is the simplest part.
All of the above AI solutions hinge on having a stable property ID that links across systems. Listings, public records, CRM contacts, valuation history, tour assets, maintenance tickets, and transactions frequently reside in different tools. If you can’t consistently link them to the same property, you can’t build a reliable system.
A popular pattern is a “property entity spine” that holds the canonical property record with pointers to source systems. You don’t have to bring everything together on day one. You need a stable mapping and a change log.
Models fit in one interest rate regime can break down in another. Neighborhoods evolve. Building booms alter comps. Seasonality changes.
You’ll need drift monitoring in your AVM and in your demand models. You monitor errors over time, you monitor shifts at the segment level, and you monitor shifts in the input feature distribution. And you schedule retraining procedures associated with market volatility type, instead of calendar time.
The U.S. interagency AVM final rule is needed because valuation models impact mortgage decisions, and agencies want quality control standards. HUD guidance exists because algorithmic tools can produce discriminatory outcomes in access to housing and advertising. These are not just abstract things. They are signals, and if you’re serious about AI real estate, you must have audit logs, documentation, vendor oversight, and monitoring.
The NIST AI Risk Management Framework is useful here as it provides a practical perspective on the lifecycle of managing AI risks, highlighting context and the requirement to observe effects over time.
It has become a design principle that in high-stakes processes such as underwriting, screening, and pricing, a human must review. The product mission is to push humans to where they have value and let AI execute triage, extraction, and pattern identification.
The best systems are clear about whether humans must review, the model can take an action, and how overrides are logged in.
A lot of teams select AI projects based on what sounds impressive. Instead, choose items that generate data and workflow efficiencies that make the following iterations simpler.
A typical high-leverage sequence is: start with lead capture and CRM hygiene (because it generates labeled data quickly), then deploy lead scoring and routing, then bring in tour/media analytics as intent signals, then create valuation enrichment, then bring in transaction document extraction, and ultimately advance into the more regulated decision tiers like screening and automated valuation for lending, based on your business design.
If you’re a marketplace, trust and fraud detection are essential because they protect the platform. If you are a property manager, maintenance and support automation is essential because it reduces cost-to-serve quickly.
Since your larger context is delivery team evaluation, here is the practical lens: You aren’t buying code. You’re purchasing the judgment.
A good team will be able to tell you in layman’s terms what information you need, what tags you will have to use, what “good” looks like operationally, and what risks you face in your corner of the world. They should talk about audit logs, monitoring, and privacy without you having to prompt them. They should be clear about the areas where AI is not dependable and where it needs to be reviewed by a human.
If a group skips straight to model architecture without considering data lineage, workflow integration, and governance, you should be wary. Real estate AI is rife with teams that can demo but can’t run.
A good partner will also guide you through the build vs. buy decisions. For example, you may purchase a 3D tour platform and integrate it, but develop your lead scoring and conversion intelligence because that’s your point of differentiation. Alternatively, you can choose a document extraction vendor, but you will need to build your validation logic and workflow around that.
